AI & Tech··1 sources verified
Training-Free Method Routes Queries to LoRA Adapters Without Retraining
Summarised by Relevant News AI · Read time: 3 min
Researchers propose LoRAWeaver, a training-free method that selects task-specific LoRA adapters by matching adapter-conditioned query representations against task prototypes using cosine similarity. Tested on LLaMA-2-7B and 13B across a 48-task FLAN-v2 benchmark and out-of-distribution tasks, it beats other training-free baselines in cross-task settings and approaches a trained retrieval encoder. An ablation attributes the gains to directional alignment in the adapter-output space.
Why it matters: Teams that accumulate many fine-tuned LoRA adapters can route requests to the right one without retraining a router every time a new adapter is added.
All sources